Skip to content
AI INTEGRATION & AUTOMATION

Add AI to what you already run.

You don't always need a new product. Often the bigger win is AI inside the app, inbox or workflow your team already uses.

01 PROBLEMS

Problems we solve

Where most of these projects start.

  1. 01

    Copying data by hand

    Your team copies details from emails and forms into other tools, one field at a time.

  2. 02

    Documents waiting in a queue

    Invoices and forms pile up until someone has time to read and enter them.

  3. 03

    AI features without a rewrite

    You want search, summaries or drafting in your product, but a rebuild isn't an option.

  4. 04

    Answers stuck with a few people

    The same questions go to the same people, because the answers live in their heads and old docs.

02 BUILD

What we build

AI features in existing apps

search, summaries, drafting and classification.

Workflow automation

the manual steps between your tools, handled automatically.

Document processing

extract data from invoices, forms and PDFs into your systems.

Assistants for your team

answer questions from your internal knowledge.

Modernisation

restructure legacy code so AI features can be added safely.

03 DELIVERABLES

What you get

Everything lands in your accounts. Nothing stays locked with us.

  • Changes merged into your repository, reviewed like any other code
  • A written baseline and target for the step we're improving
  • An evaluation set for each AI feature
  • Prompts and model settings kept in version control
  • A fallback path when the model is unsure or unavailable
  • Usage and cost tracking for model calls
  • Tests that run with the rest of your test suite
  • Docs on how each feature works and how to adjust it
04 EXAMPLE

A typical project

An example scope, not a client project.

The brief

Invoice processing that reads supplier PDFs from an inbox and posts the data to your accounting system.

In scope

  • Read invoices from a shared inbox
  • Extract supplier, amounts, tax and line items
  • Check totals and flag mismatches for review
  • Post approved invoices to your accounting system
  • A review screen for anything flagged

Milestones

  1. Baseline

    The target is agreed, and a test set is built from your past invoices.

  2. Extraction

    Upload an invoice and see the fields it reads, checked against the test set.

  3. Review flow

    Flagged invoices land in a review screen where your team approves or corrects them.

  4. Live

    Approved invoices post to your accounting system, with usage and cost tracking.

05 STACK

Tools we work with

The tools we know well for this work. The spec names the ones your project uses, and why.

Our default picks 40 tools · chosen per project in the spec
  • Languages

    • Python (default pick)
    • TypeScript (default pick)
  • Model providers

    • OpenAI (default pick)
    • Anthropic (default pick)
    • Google Gemini
    • AWS Bedrock
    • Azure OpenAI
    • Open-weight models (default pick)
  • Documents

    • Docling
    • Unstructured
    • Tesseract
    • PaddleOCR
    • AWS Textract
    • Azure Document Intelligence
  • Speech

    • Whisper
    • Deepgram
    • ElevenLabs
  • Retrieval

    • pgvector
    • Qdrant
    • LlamaIndex
  • Automation

    • n8n
    • Zapier
    • Make
    • Temporal
    • Celery
  • Business apps

    • Google Workspace
    • Microsoft 365
    • Slack
    • WhatsApp Business
    • HubSpot
    • Salesforce
    • Zoho
    • Shopify
  • Data

    • PostgreSQL (default pick)
    • DuckDB
    • pandas
    • Airbyte
  • Monitoring

    • Langfuse
    • Promptfoo
    • Sentry
FAQ

Common questions

Do we need to change our existing software?

Usually very little. We add features inside your current app or connect to it through its API, and avoid rewrites where we can.

What if the AI gets something wrong?

Low-confidence results go to a person for review, and each feature is tested on your real examples before it goes live.

Do you always use AI?

No. If a rule or a simple script solves the problem, we build that instead. It is cheaper to run and easier to maintain.

Will our data be used to train AI models?

We choose provider settings that exclude your data from training, or use open-weight models on your own servers. Discovery records which.

Have something to build?

Tell us what you're working on. We reply within one business day.